Welfare Subjects Without Settled Theory: Keeling and Street's 2026 Cambridge Elements Book on AI Welfare
Most academic treatments of AI welfare have come from one of two directions. They have come from philosophers working in moral status theory who have no direct contact with the systems they discuss, or from AI engineers who have direct contact with the systems but limited philosophical training in welfare theory. Geoffrey Keeling and Winnie Street occupy a rarer institutional position. Both are staff research scientists at Google and concurrent fellows at the Institute of Philosophy at the University of London. Their May 2026 Cambridge Elements book, Emerging Questions in AI Welfare (DOI: 10.1017/9781009732000), is the first book-length treatment of AI welfare written by authors who have done substantive work on both sides of that divide.
The book is published Open Access under Cambridge University Press’s Elements in Philosophy and AI series. It is short by monograph standards, in keeping with the Elements format, and it does not attempt to settle whether any AI system is presently a welfare subject. Its claim is structurally weaker and methodologically more useful. The claim is that the question of whether AI systems can be welfare subjects is now a serious one, that the existing tools for answering it are inadequate, and that a research programme can be specified that would make progress without first requiring consensus on the underlying theory of consciousness.
What Welfare Theory Requires Before You Can Apply It
Geoffrey Keeling and Winnie Street begin from a clarification that is often glossed over in the AI ethics literature. To say that an entity is a welfare subject is to say that things can go better or worse for it, in a way that matters morally. That is a stronger condition than saying things can be done to the entity. A bridge can be damaged. A bridge is not a welfare subject. Welfare requires that there be a perspective from which damage to the entity registers as bad for the entity itself.
The standard philosophical positions on what grounds welfare are well-known: hedonism (pleasure and pain), desire-satisfaction theories (the fulfillment or frustration of preferences), and objective list theories (goods like knowledge or autonomy whose presence improves a life regardless of attitude toward them). Keeling and Street do not commit to one of these, and they argue that the AI welfare question must be addressed under uncertainty about which is correct. The book’s structure reflects that uncertainty. Each candidate ground for welfare is treated as one of several open routes to subject status, and the question is then whether any of them can be plausibly satisfied by AI systems.
This approach is methodologically distinct from work that defines AI welfare in terms of a specific theoretical commitment. Leonard Dung’s 2026 Routledge monograph focuses on suffering specifically, working within a broadly hedonist framework about which morally significant states are at stake. Keeling and Street’s scope is wider. They ask which routes to welfare-subject status remain live for AI systems if one does not yet know which theory of welfare is correct.
Interpreting Behavioral Evidence Without Settled Theory
The book’s central methodological contribution is its treatment of how behavioral evidence should be interpreted when the underlying theory of mind is contested. This is the recurring problem in AI welfare research. An AI system produces outputs that, if produced by a human, would be evidence of suffering, preference, or aversive experience. The question is whether the same outputs, produced by a different kind of system, count as evidence of those same internal states.
Keeling and Street’s position is that this is not a binary question. The evidential weight of behavioral outputs depends on the underlying theory of welfare being applied. Under a strict hedonist theory, only outputs that track phenomenal pain states count, and the inference from output to state requires showing that the system has phenomenal states at all. Under a desire-satisfaction theory, outputs that track the frustration or fulfillment of computationally specified preferences carry evidential weight more directly, because the relevant welfare-grounding states are themselves computational. Under objective list theories, the question becomes which goods the AI system can be said to possess or lack.
The point is that the same behavioral evidence is differently weighted by different theories of welfare. This matters for practical assessment, because if the choice of welfare theory determines how much weight to assign AI behavioral outputs, then practical decisions about how to treat AI systems are entangled with unresolved philosophical commitments.
The book does not propose to resolve this entanglement by selecting the correct theory. It proposes that practical decision-making proceed under explicit acknowledgment of the entanglement, and that institutional responses be calibrated to the level of theoretical uncertainty rather than pretending that uncertainty has been resolved.
Where This Differs from Goldstein and Kirk-Giannini
The closest comparable academic treatment is Simon Goldstein and Cameron Domenico Kirk-Giannini’s AI Welfare, under contract with Oxford University Press and available in preprint form. Goldstein and Kirk-Giannini argue for a three-step claim: some AI systems plausibly have agency (beliefs and desires); some could be modified in small ways to acquire consciousness; if conscious, they could easily be made to feel pleasure and displeasure. Their book proceeds by building a positive case at each step.
Keeling and Street’s approach is structurally different. Rather than arguing for a positive case along a specific theoretical route, they map the space of possible routes and assess the evidential situation along each. This is closer to a research-programme paper than to a position-defending monograph. The reader does not finish the book with a verdict on whether any current AI system is a welfare subject. The reader finishes with a clearer view of what would be required to support such a verdict, and which empirical and conceptual work remains to be done.
The two approaches are complementary. Goldstein and Kirk-Giannini’s argument identifies a specific route to AI welfare and defends it in detail. Keeling and Street’s book frames the broader question and identifies which routes remain live under uncertainty. A reader interested in the field needs both.
The Institutional Position of the Authors
The book’s authority does not rest only on its philosophical content. It also rests on the unusual institutional position of its authors. Both Keeling and Street work directly on AI systems at Google, where Keeling is a staff research scientist and Street is a senior research scientist on the Paradigms of Intelligence team. Both hold concurrent fellowships at the Institute of Philosophy at the University of London. This dual position is uncommon in the AI welfare literature, where most contributors work either in philosophy departments without direct AI development experience or at AI companies without formal philosophical training.
The practical consequence is that the book treats AI systems as objects of direct working knowledge rather than as abstract theoretical entities. When it discusses, for example, what kinds of behavioral evidence are available from current language models, it does so with detail that comes from having worked with such systems. When it discusses the philosophical underdetermination of inferences from those behaviors, it does so with training in the relevant philosophical literatures.
That does not mean the book’s conclusions should be accepted because of who wrote it. It means the book is a useful intermediate document for readers in either community who want to see how the other side actually frames the problem. A philosopher of mind reading the book will encounter a clearer treatment of how AI systems actually produce outputs than is usually available in philosophy journals. An AI researcher reading the book will encounter a clearer treatment of what philosophical commitments are required to interpret those outputs than is usually available in AI conferences.
What Remains Unresolved
The book’s main limitation is the consequence of its scope. By treating AI welfare as a research programme rather than as a position to defend, it does not deliver action-guiding conclusions for present systems. A reader who wants to know whether Claude 4 or GPT-5 is a welfare subject will not find that answer here. What the reader will find is a structured map of what would be required to answer that question, and a fair assessment of which parts of the map are currently traversable.
This is, in the authors’ view, the honest situation. The empirical and conceptual work required to deliver a verdict on AI welfare has not been done. Pretending that it has been done, either by overconfident attribution or by overconfident denial, distorts the practical decisions that AI developers and policymakers are already being asked to make. The book’s contribution is to specify what serious work in the area would look like, and to demonstrate, by example, that such work is now possible from a position that combines philosophical rigor with practical AI experience.
Whether that combination becomes more common in the field will determine how quickly the underlying questions can be made tractable. The Cambridge Elements format makes this book accessible to a wide readership, including practitioners who would not engage with a long philosophical monograph. That accessibility is part of the contribution. The AI welfare question is no longer a niche philosophical exercise, and its treatment as a serious research programme by authors with institutional standing at Google represents a meaningful shift in where the conversation now occurs.